AUTHOR=Zuo Teng , Zheng Yanhua , He Lingfeng , Chen Tao , Zheng Bin , Zheng Song , You Jinghang , Li Xiaoyan , Liu Rong , Bai Junjie , Si Shuxin , Wang Yingying , Zhang Shuyi , Wang Lili , Chen Jianhui TITLE=Automated Classification of Papillary Renal Cell Carcinoma and Chromophobe Renal Cell Carcinoma Based on a Small Computed Tomography Imaging Dataset Using Deep Learning JOURNAL=Frontiers in Oncology VOLUME=11 YEAR=2021 URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2021.746750 DOI=10.3389/fonc.2021.746750 ISSN=2234-943X ABSTRACT=Objectives

This study was conducted in order to design and develop a framework utilizing deep learning (DL) to differentiate papillary renal cell carcinoma (PRCC) from chromophobe renal cell carcinoma (ChRCC) using convolutional neural networks (CNNs) on a small set of computed tomography (CT) images and provide a feasible method that can be applied to light devices.

Methods

Training and validation datasets were established based on radiological, clinical, and pathological data exported from the radiology, urology, and pathology departments. As the gold standard, reports were reviewed to determine the pathological subtype. Six CNN-based models were trained and validated to differentiate the two subtypes. A special test dataset generated with six new cases and four cases from The Cancer Imaging Archive (TCIA) was applied to validate the efficiency of the best model and of the manual processing by abdominal radiologists. Objective evaluation indexes [accuracy, sensitivity, specificity, receiver operating characteristic (ROC) curve, and area under the curve (AUC)] were calculated to assess model performance.

Results

The CT image sequences of 70 patients were segmented and validated by two experienced abdominal radiologists. The best model achieved 96.8640% accuracy (99.3794% sensitivity and 94.0271% specificity) in the validation set and 100% (case accuracy) and 93.3333% (image accuracy) in the test set. The manual classification achieved 85% accuracy (100% sensitivity and 70% specificity) in the test set.

Conclusions

This framework demonstrates that DL models could help reliably predict the subtypes of PRCC and ChRCC.